<h1 id="description">Description</h1>Based on my extensive testing with SDXL anime-style models, I found that most base models stubbornly refuse to depict characters with "tanned skin" or "muscles" as the "Bottom (Receiver)" in BL/Yaoi contexts.While individual character LoRAs for Cú Chulainn and Emiya already exist, this specific LoRA focuses entirely on their relationship dynamics. It is technically impossible to completely "lock" roles using only a LoRA, but this model drastically improves the success rate, which is otherwise extremely low without it. The dataset is based on my personal art style, regenerated and refined to capture a specific aesthetic.<ul><li>As of now, the Pony version is superior in terms of prompt flexibility and situational variety. However, for clearly defined scenes, Illustrious may produce better results when given precise, well-crafted prompts.</li><li>今処、lycoris(単独用よりはloraの性能サポート用)を制作中です。Please read the following guidelines before use:</li></ul><h3 id="1.-dataset-disclaimer">1. Dataset Disclaimer</h3>Since the training dataset includes AI-regenerated images based on my original sketches/drawings, I do not claim the training data as 100% my own hand-drawn art. However, it clearly reflects my specific taste and drawing style.<h3 id="2.-permissions">2. Permissions</h3>You are free to merge this with other LoRAs or Checkpoints. However, re-uploading (reposting) this raw file is strictly prohibited.<h3 id="3.-recommended-settings">3. Recommended Settings</h3>This model is a calibrated merge of three distinct LoRAs (Logic, Texture, and Atmosphere).<s>Recommended Weight: 1.0</s>It depends on model and situation. 0.6~1.0 is sweet spot. pony is OK when stronger CFG, but illustrious needs more gentle SFG. Note: Since the optimal ratio is already baked in, lowering the weight significantly may cause the role-locking logic to fail.<h3 id="4.-role">4. Role</h3>Locking Limitations As mentioned, a LoRA alone cannot guarantee a 100% role fix. If you require absolute precision in positioning (e.g., specific poses), I highly recommend using Regional ControlNet alongside this model.<h3 id="5.-prompting-tip-(crucial)-for-archer's-skin-tone:">5. Prompting Tip (Crucial) For Archer's skin tone:</h3>Avoid: dark skin (This tag often bleeds onto Lancer or fails to apply correctly).Use: tanned skin (This yields the most accurate results).<h3 id="6.-usage-and-quality-note">6. Usage &amp; Quality Note</h3>Although explicit (XXX-rated) samples are not displayed here, this LoRA was fundamentally designed with NSFW generation in mind. It is capable of generating general slice-of-life scenarios, but please understand that the quality for broad SFW situations might be slightly lower compared to NSFW outputs. I will consider improving SFW stability in future updates.(Optional) Recommended Tags: yrym, yariyumi, tanned skin, muscular

Ghost & Darkness: Fate Lancer x Archer (Role-Fixed / No Swap)

<pre><code>⚠️ URGENT NOTICE: Without pairing this with other LoRAs, character bleeding issues persist in CivitAI online generation. For now, I recommend using this LoRA for img2img only. I will push an updated version as soon as the issue is resolved to a satisfactory degree.</code></pre><pre><code></code></pre><pre><code></code></pre>GnD was originally built for Illustrious, and that will remain the main line going forward — because the models that best understand the relational dynamics between characters, which is the very foundation of shipping/coupling, have always been in the Illustrious family. (GnD is fundamentally created by testing and merging LoRAs trained on models that excel at capturing those dynamics.)That said, I've always loved the synergy between Pony's distinctive color palette and art style with my ship. I considered merging the models themselves, but building an entire checkpoint solely for one coupling felt like too massive an undertaking.This might partly be because I'm not deeply embedded in Reddit communities, but from what I've seen, most LoRA training in the community focuses on individual characters — or even when multiple characters are involved, the training rarely explores how those characters relate to each other. That approach doesn't align well with teaching the concept of a ship/pairing.For example, clip skip is typically set to 2 for character training, but in my case, setting clip skip to 1 (CivitAI standard) almost always yielded better results.So this version was trained with a dataset built specifically and thoroughly for Pony — as a dedicated Pony edition. Since Illustrious LoRAs generally don't transfer well to Pony, this dataset is completely different from Version 1, designed from the ground up for Pony.With that context out of the way, please keep the following in mind when using this LoRA:1) Dataset &amp; Copyright This LoRA is trained on my own artwork — including originals, AI hi-res upscales, and lightly rearranged versions of my drawings. However, through extensive experimentation (enough to reach Top Trainer #15 on CivitAI from online training alone, not counting local runs), I've come to understand that the composition, lighting, and overall output can change drastically depending on which model you generate from. For that reason, I will not assert copyright claims over the dataset.2) Character Roles &amp; Appearance This LoRA is tuned to mostly maintain Lancer as top and Archer as bottom, along with each character's distinct appearance. However, due to inherent SDXL tendencies — which are even more pronounced in Pony — 100% control is not possible.3) Usage &amp; Redistribution You are free to mix, merge, or combine this LoRA with other LoRAs or checkpoints for personal use. However, redistribution of the original or any merged version is strictly prohibited.4) Archer's Skin Tone It's safest to specify Archer's skin as <code>tanned_skin</code>.5) Tag Order Matters — A Lot Pony is extremely sensitive to tag order. As a general rule, list Lancer's traits first, then Archer's. However, if the composition has one character barely visible, or if Archer is the focal "lead" in terms of camera framing, putting Archer's tags first may produce better results. This is something you'll need to fine-tune by reviewing your outputs.6) Recommended Weight &amp; Complementary LoRAs Best results at around 0.6–0.8 weight. I strongly recommend pairing this with a Pony-specific detailer or anatomy correction LoRA for optimal output.

Files with SHA256: 6bcf9d32e479bf26a23372d7c26c2fc583ab42b41fa9433af72f40ae95a6e7b6